Target-Dependent Performance of Molecular Similarity Metrics in Drug Discovery

ABSTRACT

Pharmaceutical companies screen millions of compounds computationally before running experiments, relying on the molecular similarity principle that structurally similar molecules exhibit similar biological activities. However, whether this principle works equally well across different protein types remains unclear. This study evaluated eight therapeutically relevant proteins spanning five families—kinases, G-protein coupled receptors, proteases, metalloenzymes, and nuclear receptors—using standardized datasets of 2,000 compounds per target. Four molecular fingerprint methods were compared to quantify the similarity-activity relationship. Results revealed target-dependent variation. CDK2 kinase showed strong correlation between similarity and activity (r = -0.191), while thrombin protease showed essentially none (r = -0.006). Kinases consistently demonstrated robust correlations (mean r = -0.147), whereas proteases showed weak correlations (mean r = -0.023), representing a sevenfold difference. Activity cliffs—structurally similar molecules with dramatically different activities—occurred at 6.2% overall but varied threefold across families. Morgan radius 2 fingerprints performed best on average, though target-specific selection improved accuracy by 68%. These findings demonstrate that binding site architecture determines similarity principle and provides practical guidance for selecting screening methods in drug discovery.

INTRODUCTION.

Developing new drugs costs over $1.5 billion per approved medicine and requires 10–15 years [1]. To reduce these costs, pharmaceutical companies use computational screening to predict which molecules might work before synthesizing and testing them experimentally. This approach depends on the molecular similarity principle: structurally similar molecules should exhibit similar biological activities [2]. The practical application is straightforward. When a chemist discovers that molecule A inhibits a cancer-causing protein, he searches databases for molecules structurally similar to A. The expectation is that these similar molecules will also inhibit the target, allowing researchers to focus on perhaps 100 most promising drug candidates instead of testing millions of molecules randomly [3].

Despite widespread use of this principle, a critical question remains unresolved: when does similarity-based screening actually work? Martin and colleagues raised this concern in 2002, asking whether structurally similar molecules truly have similar biological activity [4]. However, medicinal chemists still lack clear guidelines. Thus, a pharmaceutical company might invest thousands of dollars synthesizing 50 compounds identified through similarity screening, only to discover that none work because similarity is uninformative for that protein.

Molecular structures are encoded computationally using fingerprints, which are mathematical representations that convert complex three-dimensional molecules into sets of numbers that computers can compare [5]. Several fingerprint types exist, each capturing different chemical intuitions about which structural features matter for biological activity [6]. The challenge facing medicinal chemists is: which fingerprint should be used when screening against a new protein target? Most researchers currently select one arbitrarily or accept whatever their software defaults to. This arbitrary choice could prove costly if a different fingerprint would work better for their target.

Another complication arises from activity cliffs, which are molecule pairs that appear nearly identical structurally but behave completely different biologically [7-9]. These activity cliffs represent direct violations of the similarity principle. Understanding when activity cliffs occur would help researchers identify situations where similarity-based screening is unreliable.

This study addresses a critical question: do different protein families follow the similarity principle to different degrees? The hypothesis tested was that kinases should show stronger similarity-activity relationships than proteases because kinase binding pockets are geometrically constrained [10, 11]. A kinase’s active site forms a well-defined pocket that similar molecules should occupy similarly. In contrast, a protease’s active site contains multiple flexible subsites that similar molecules might occupy differently, potentially breaking the similarity-activity relationship [12].

METHODS.

Target Selection and Data Collection.

Eight therapeutically relevant proteins were selected spanning five major drug target families as shown in Table 1 [10, 12-14]. These proteins collectively address major disease areas including cancer, psychiatric disorders, cardiovascular disease, and metabolic disorders. Bioactivity data was obtained from ChEMBL, a freely available database containing millions of measurements documenting how strongly small molecules bind to different proteins [15].

Table 1. The list of eight protein targets studied in this work.
Target	Family	Disease Area	CHEMBL codes
CDK2	Kinase	Cancer	CHEMBL1862
EGFR	Kinase	Cancer	CHEMBL203
Dopamine D2	GPCR	Psychiatric	CHEMBL217
Adenosine A2a	GPCR	Parkinson’s	CHEMBL220
Cannabinoid CB1	GPCR	Pain	CHEMBL1827
Thrombin	Protease	Blood clots	CHEMBL204
Carbonic anhydrase II	Metalloenzyme	Glaucoma	CHEMBL1871
Androgen receptor	Nuclear receptor	Cancer	CHEMBL233

For each target, compounds meeting strict criteria were selected: valid chemical structures, molecular weight 150–650 Daltons, activity values (pChEMBL, which equals -log₁₀ of IC₅₀ in molar units) between 4.0 and 11.0, and no statistical outliers beyond three standard deviations. When the same molecule appeared multiple times, only the highest potency measurement was retained. Following this rigorous filtering, exactly 2,000 compounds per target were selected through stratified random sampling. Activity values were divided into five equal-frequency bins, and compounds were randomly sampled proportionally from each bin. This stratification ensures that all activity levels from weak to strong are represented proportionally.

Fingerprint Generation.

Four fingerprint types were generated using RDKit software: Morgan radius 2 (1024 bits) captures circular neighborhoods extending two bonds from each atom; Morgan radius 3 (1024 bits) uses larger neighborhoods capturing more context; MACCS keys (166 bits) encode predefined structural patterns; RDKit topological (2048 bits) hashes all paths through the molecule up to seven bonds [5, 6, 16].

Similarity Analysis.

For each protein target and fingerprint type, similarity was calculated between every possible molecule pair using the Tanimoto coefficient, which ranges from 0 for no shared features to 1 for identical structures [17]. With 2,000 molecules per target, this produces millions of unique pairs. For each pair, two quantities were computed: structural similarity represented by Tanimoto coefficient and activity difference represented by absolute difference in pChEMBL values. We calculated the Pearson correlation between these quantities across all pairs [18]. This correlation quantifies whether high structural similarity corresponds with small activity differences. A correlation of -1 would indicate perfect predictive power, while 0 would indicate that similarity provides no information about activity.

Activity Cliff Detection.

Activity cliffs were identified using standard medicinal chemistry definitions: a molecule pair qualifies if it exhibits Tanimoto similarity ≥ 0.85 and activity difference ≥ 2.0 pChEMBL units. The cliff rate was calculated as the percentage of highly similar pairs showing large activity differences, directly measuring how often similarity-based predictions fail.

RESULTS.

Kinases Follow the Similarity Principle While Proteases Do Not.

We observed that molecular similarity predicts biological activity for some proteins but fails for others. Figure 1 shows correlation between structural similarity and activity difference for all eight targets using four fingerprint types. CDK2 kinase showed the strongest relationship, achieving Pearson correlation r = -0.191 using Morgan radius 2 fingerprints. In contrast, thrombin protease showed essentially no relationship, with r = -0.006 using MACCS keys. This represents a 32-fold difference in correlation magnitude. Thrombin’s near-zero correlation implies that if two molecules share 90% Tanimoto similarity for thrombin, this provides zero information about whether they differ in activity by 10-fold or 1,000-fold. Structural similarity measurements are effectively worthless for predicting thrombin activity.

The remaining six targets spanned the spectrum between these extremes. Cannabinoid CB1 receptor ranked second with r = -0.155 using RDKit fingerprints. EGFR kinase achieved r = -0.125 with Morgan radius 2. Androgen receptor and carbonic anhydrase II showed moderate correlations around r = -0.11 to -0.13. Dopamine D2 receptor showed only r = -0.062, and adenosine A2a receptor achieved merely r = -0.064.

CDK2 showed nearly identical correlations across different fingerprint types. Morgan radius 2 yielded r = -0.191, Morgan radius 3 yielded r = -0.174, and RDKit topological yielded r = -0.173. These three fingerprints encode structure using different algorithms, yet all independently converged on the same conclusion. This consistency suggests detection of a biological signal rather than noise in data.

Protein Families Show Systematic Differences.

Grouping targets by protein family revealed clear patterns (Figure 2). Kinases performed best, averaging r = -0.147 across all fingerprint types. The consistency between CDK2 (serine/threonine kinase) and EGFR (tyrosine kinase) suggests that properties of ATP-binding pockets make them inherently suitable for similarity-based approaches.

GPCRs showed internal variation: mean correlation r = -0.094 with standard deviation 0.042. Cannabinoid CB1 receptor followed the similarity principle robustly (r = -0.133 average), while dopamine D2 (r = -0.044) and adenosine A2a (r = -0.040) barely followed it. This heterogeneity indicates that GPCR family membership alone does not determine whether similarity works. Instead, individual binding site characteristics matter more.

Serine proteases showed the weakest performance (mean r = -0.023), strongly supporting the hypothesis. The sevenfold difference between kinases (r = -0.147) and proteases (r = -0.023) has direct practical implications for drug discovery. Nuclear receptors (androgen receptor, mean r = -0.086) and metalloenzymes (carbonic anhydrase II, mean r = -0.082) showed intermediate performance.

Target-Specific Fingerprint Selection Improves Performance.

Averaged across all eight targets, Morgan radius 2 fingerprints achieved best overall performance with mean correlation r = -0.106 (Figure 3). This outperformed Morgan radius 3 (r = -0.097), RDKit topological and MACCS keys. However, important target-specific patterns emerged. For CDK2, all fingerprint types performed similarly well. For EGFR and cannabinoid CB1, RDKit fingerprints outperformed Morgan fingerprints. For carbonic anhydrase II and adenosine A2a, Morgan radius 3 showed advantages.

Implementing target-specific fingerprint selection, or choosing the best-performing fingerprint for each individual target, improved the mean absolute correlation from -0.106 to -0.178, representing a 68% improvement. This enhancement justifies conducting small-scale validation studies before launching large screening campaigns.

Activity Cliffs.

Across all targets and fingerprints, roughly one in sixteen highly similar pairs showed large activity differences, limiting the prediction accuracy, and representing an overall 6.2% cliff rate (Figure 4). Rates varied by family. Nuclear receptors showed high rates, while Ser/Thr kinases showed very low rates (2.8%). The inverse relationship between cliff rate and correlation strength makes biological sense. Activity cliffs are violations of the similarity principle, so their prevalence determines how often similarity-based predictions fail.

Finally, in Figure 5, molecule pairs demonstrating activity cliffs are shown. Each pair displays the activity difference (Δ), similarity score (Sim), and whether the molecules share the same scaffold. Two of the three cliffs involve slightly different scaffolds, suggesting that structurally distinct cores can appear similar to 2D fingerprints. The one same-scaffold cliff (Cliff 3) identifies a structural hot spot where minor substituent changes cause dramatic activity changes. Such examples illustrate cases where similarity-based predictions fail.

DISCUSSION.

The hypothesis that kinases would demonstrate stronger adherence to the similarity principle than proteases received strong support. The sevenfold difference in correlation magnitude (kinases r = -0.147 versus proteases r = -0.023) reflects differences in binding site architecture. Kinase ATP-binding sites are geometrically constrained [10]. The hinge region forms conserved hydrogen bonds. Similar two-dimensional structures tend to adopt similar three-dimensional binding poses in this constrained environment [19]. Molecules sharing structural features therefore yield similar affinities.

Thrombin presents a different scenario. The active site contains multiple distinct subsites (S1, S2, S3, S4) that accommodate diverse functional groups flexibly [12]. Two molecules might share high two-dimensional similarity yet differ dramatically in subsite occupancy. These differences affect binding affinity despite high two-dimensional similarity [20].

GPCR heterogeneity likely reflects binding pocket diversity. Cannabinoid CB1 receptor has an enclosed binding cavity [13]. This geometric constraint favors the similarity principle because ligands must fit within confined space. Adenosine A2a receptor has a more open pocket, allowing diverse ligands to satisfy binding requirements [14].

These findings carry practical implications. Before launching large-scale virtual screening campaigns, researchers should conduct small validation studies with 50–100 known active compounds to test fingerprint performance for their target. This requires only hours of computational time but reveals whether similarity searching will work. For kinases, the similarity principle is reliable. For GPCRs, performance varies, thus needing to validate the specific target. For proteases, do not rely on two-dimensional similarity alone, and instead combine with structure-based methods like molecular docking. The 6.2% overall activity cliff rate aligns with previous studies [9]. Activity cliffs are not rare anomalies but common features that limit prediction accuracy. Same-scaffold activity cliffs identify positions where small changes cause large effects, deserving attention.

In conclusion, the molecular similarity principle’s validity depends critically on target protein binding site architecture. Kinases show robust adherence while proteases show weak adherence, with GPCRs showing target-specific variation. Morgan radius 2 fingerprints perform best on average, but target-specific selection substantially improves results. These findings provide guidance for similarity method selection in drug discovery, potentially saving time and resources by identifying which approaches will succeed for specific protein targets.

ACKNOWLEDGMENTS.

The ChEMBL team is acknowledged for maintaining the publicly accessible bioactivity database. The RDKit developers and open-source Python community enabled this computational work.

CODE AVAILABILITY.

The full python code is available at:

github.com/apoorvapu/data_science/blob/main/moleculeActivitySimilarity.ipynb

REFERENCES.

1. Bajorath, in Bioinformatics: Volume II: Structure, Function, and Applications. (Springer, 2016), pp. 231-245.
2. B. Hendrickson, Concepts and Applications of Molecular Similarity. Science 252, 1189-1190 (1991).
3. C. D. Hawkins, A. G. Skillman, A. Nicholls, Comparison of shape-matching and docking as virtual screening tools. Journal of Medicinal Chemistry 50, 74-82 (2007).
4. C. Martin, J. L. Kofron, L. M. Traphagen, Do structurally similar molecules have similar biological activity? Journal of Medicinal Chemistry 45, 4350-4358 (2002).
5. Willett, Similarity-based virtual screening using 2D fingerprints. Drug Discovery Today 11, 1046-1053 (2006).
6. Riniker, G. A. Landrum, Open-source platform to benchmark fingerprints for ligand-based virtual screening. Journal of Cheminformatics 5, 26 (2013).
7. Bajorath, Representation and identification of activity cliffs. Expert Opinion on Drug Discovery 12, 879-883 (2017).
8. Stumpfe, J. Bajorath, Exploring activity cliffs in medicinal chemistry: miniperspective. Journal of Medicinal Chemistry 55, 2932-2942 (2012).
9. M. Maggiora. (ACS Publications, 2006), vol. 46, pp. 1535-1535.
10. Jia, C. M. Quinn, A. I. Gagnon, R. Talanian, Homogeneous time-resolved fluorescence and its applications for kinase assays in drug discovery. Analytical biochemistry 356, 273-281 (2006).
11. Roskoski Jr., Classification of small molecule protein kinase inhibitors based upon the structures of their drug-enzyme complexes. Pharmacological Research 103, 26-48 (2016).
12. A. Huntington, Molecular recognition mechanisms of thrombin. Journal of Thrombosis and Haemostasis 3, 1861-1872 (2005).
13. Hua et al., Crystal structure of the human cannabinoid receptor CB1. Cell 167, 750-762 (2016).
14. -P. Jaakola et al., The 2.6 angstrom crystal structure of a human A2A adenosine receptor bound to an antagonist. Science 322, 1211-1217 (2008).
15. Gaulton et al., The ChEMBL database in 2017. Nucleic Acids Research 45, D945-D954 (2017).
16. Rogers, M. Hahn, Extended-connectivity fingerprints. Journal of Chemical Information and Modeling 50, 742-754 (2010).
17. Bajusz, A. Rácz, K. Héberger, Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations? Journal of Cheminformatics 7, 20 (2015).
18. Pearson, VII. Note on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London 58, 240-242 (1895).
19. Martin, P. Mukherjee, D. Sullivan, J. Jansen, Profile-QSAR: a novel meta-QSAR method that combines activities across the kinase family to accurately predict affinity, selectivity, and cellular activity. Journal of Chemical Information and Modeling 51, 1942-1956 (2011).
20. Bode, D. Turk, A. Karshikov, The refined 1.9‐Å X‐ray crystal structure of d‐Phe‐Pro‐Arg chloromethylketone‐inhibited human α‐thrombin: Structure analysis, overall structure, electrostatic properties, detailed active‐site geometry, and structure‐function relationships. Protein Science 1, 426-471 (1992).

---

---

Posted by buchanle on Friday, June 5, 2026 in May 2026.

Tags: activity cliffs, fingerprints, molecular similarity, structure-activity relationships, virtual screening